NLE: Non-autoregressive LLM-based ASR by Transcript Editing

Benchmark Model Rank Results
speech-recognition-on-amiibm-granite/granite-speech-4.1-2b-nar#1Word Error Rate (WER): 7.64
speech-recognition-on-ami-cleanedibm-granite/granite-speech-4.1-2b-nar#1Word Error Rate (WER): 6.96
speech-recognition-on-earnings-22ibm-granite/granite-speech-4.1-2b-nar#1Word Error Rate (WER): 8.15
speech-recognition-on-gigaspeech-cleanedibm-granite/granite-speech-4.1-2b-nar#30Word Error Rate (WER): 8.62
speech-recognition-on-gigaspeech-testibm-granite/granite-speech-4.1-2b-nar#28Word Error Rate (WER): 8.67
speech-recognition-on-librispeech-test-cleanibm-granite/granite-speech-4.1-2b-nar#4Word Error Rate (WER): 1.04
speech-recognition-on-librispeech-test-otheribm-granite/granite-speech-4.1-2b-nar#6Word Error Rate (WER): 2.4
speech-recognition-on-spgispeechibm-granite/granite-speech-4.1-2b-nar#37Word Error Rate (WER): 3.23
speech-recognition-on-voxpopuliibm-granite/granite-speech-4.1-2b-nar#11Word Error Rate (WER): 5.63
speech-recognition-on-voxpopuli-cleanedibm-granite/granite-speech-4.1-2b-nar#41Word Error Rate (WER): 4.25